Information Retrieval: Beyond Ranking of Pages
21.8 Information Retrieval: Beyond Ranking of Pages
Information-retrieval systems were originally designed to find textual documents related to a query, and later extended to finding pages on the Web that are related to a query. People use search engines for many different tasks, from simple tasks such as locating a Web site that they want to use, to a broader goal of finding information on a topic of interest. Web search engines have become extremely good at the task of locating Web sites that a user wants to visit. The task of providing information on a topic of interest is much harder, and we study some approaches in this section.
There is also an increasing need for systems that try to understand documents (to a limited extent), and answer questions based on the (limited) understanding. One approach is to create structured information from unstructured documents
932 Chapter 21 Information Retrieval
and to answer questions based on the structured information. Another approach applies natural language techniques to find documents related to a question (phrased in natural language) and return relevant segments of the documents as an answer to the question.
21.8.1 Diversity of Query Results
Today, search engines do not just return a ranked list of Web pages relevant to
a query. They also return image and video results relevant to a query. Further, there are a variety of sites providing dynamically changing content such as sports scores, or stock market tickers. To get current information from such sites, users would have to first click on the query result. Instead, search engines have created “gadgets,” which take data from a particular domain, such as sports updates, stock prices, or weather conditions, and format them in a nice graphical manner, to be displayed as results for a query. Search engines have to rank the set of gadgets available in terms of relevance to a query, and display the most relevant gadgets, along with Web pages, images, videos, and other types of results. Thus
a query result has a diverse set of result types. Search terms are often ambiguous. For example, a query “eclipse” may be referring to a solar or lunar eclipse, or to the integrated development environment ( IDE ) called Eclipse. If all the highly ranked pages for the term “eclipse” are about the IDE , a user looking for information about solar or lunar eclipses may be very dissatisfied. Search engines therefore attempt to provide a set of results that are diverse in terms of their topics, to minimize the chance that a user would be dissatisfied. To do so, at indexing time the search engine must disambiguate the sense in which a word is used in a page; for example, it must decide whether the use of the word “eclipse” in a page refers to the IDE or the astronomical phenomenon. Then, given a query, the search engine attempts to provide results that are relevant to the most common senses in which the query words are used.
The results obtained from a Web page need to be summarized as a snippet in
a query result. Traditionally, search engines provided a few words surrounding the query keywords as a snippet that helps indicate what the page contains. However, there are many domains where the snippet can be generated in a much more meaningful manner. For example, if a user queries about a restaurant, a search engine can generate a snippet containing the restaurant’s rating, a phone number, and a link to a map, in addition to providing a link to the restaurant’s home page. Such specialized snippets are often generated for results retrieved from a database, for example, a database of restaurants.
21.8.2 Information Extraction
Information-extraction systems convert information from textual form to a more structured form. For example, a real-estate advertisement may describe attributes of a home in textual form, such as “two-bedroom three-bath house in Queens, $1 million”, from which an information extraction system may extract attributes such as number of bedrooms, number of bathrooms, cost and neighborhood. The original advertisement could have used various terms such as 2BR, or two BR,
21.8 Information Retrieval: Beyond Ranking of Pages 933
or two bed, to denote two bedrooms. The extracted information can be used to structure the data in a standard way. Thus, a user could specify that he is interested in two-bedroom houses, and a search system would be able to return all relevant houses based on the structured data, regardless of the terms used in the advertisement.
An organization that maintains a database of company information may use an information-extraction system to extract information automatically from newspaper articles; the information extracted would relate to changes in attributes of interest, such as resignations, dismissals, or appointments of company officers.
As another example, search engines designed for finding scholarly research articles, such as Citeseer and Google Scholar, crawl the Web to retrieve documents that are likely to be research articles. They examine some features of each retrieved document, such as the presence of words such as “bibliography”, “references”, and “abstract”, to judge if a document is in fact a scholarly research article. They then extract the title, list of authors, and the citations at the end of the article, by using information extraction techniques. The extracted citation information can
be used to link each article to articles that it cites, or to articles that cite it; such citation links between articles can be very useful for a researcher. Several systems have been built for information extraction for specialized ap- plications. They use linguistic techniques, page structure, and user-defined rules for specific domains such as real estate advertisements or scholarly publications. For limited domains, such as a specific Web site, it is possible for a human to spec- ify patterns that can be used to extract information. For example, on a particular Web site, a pattern such as “Price: <number> $”, where <number> indicates any number, may match locations where the price is specified. Such patterns can be created manually for a limited number of Web sites.
However, on the Web scale with millions of Web sites, manual creation of such patterns is not feasible. Machine-learning techniques, which can learn such patterns given a set of training examples, are widely used to automate the process of information extraction.
Information extraction usually has errors in some fraction of the extracted information; typically this is because some page had information in a format that syntactically matched a pattern, but did not actually specify a value (such as the price). Information extraction using simple patterns, which separately match parts of a page, is relatively error prone. Machine-learning techniques can perform much more sophisticated analysis, based on interactions between patterns, to minimize errors in the information extracted, while maximizing the amount of information extracted. See the references in the bibliographical notes for more information.
21.8.3 Question Answering
Information retrieval systems focus on finding documents relevant to a given query. However, the answer to a query may lie in just one part of a document, or in small parts of several documents. Question answering systems attempt to provide direct answers to questions posed by users. For example, a question of the
934 Chapter 21 Information Retrieval
form “Who killed Lincoln?” may best be answered by a line that says “Abraham Lincoln was shot by John Wilkes Booth in 1865.” Note that the answer does not actually contain the words “killed” or “who”, but the system infers that “who” can be answered by a name, and “killed” is related to “shot”.
Question answering systems targeted at information on the Web typically generate one or more keyword queries from a submitted question, execute the keyword queries against Web search engines, and parse returned documents to find segments of the documents that answer the question. A number of linguistic techniques and heuristics are used to generate keyword queries, and to find relevant segments from the document.
An issue in answering questions is that different documents may indicate different answers to a question. For example, if the question is “How tall is a giraffe?” different documents may give different numbers as an answer. These answers form a distribution of values, and a question answering system may choose the average, or median value of the distribution as the answer to be returned; to reflect the fact that the answer is not expected to be precise, the system may return the average along with the standard deviation (for example, average of 16 feet, with a standard deviation of 2 feet), or a range based on the average and the standard deviation (for example, between 14 and 18 feet).
Current-generation question answering systems are limited in power, since they do not really understand either the question or the documents used to answer the question. However, they are useful for a number of simple question answering tasks.
21.8.4 Querying Structured Data
Structured data are primarily represented in either relational or XML form. Several systems have been built to support keyword querying on relational and XML data (see Chapter 23). A common theme between these systems lies in finding nodes (tuples or XML elements) containing the specified keywords, and finding connecting paths (or common ancestors, in the case of XML data) between them.
For example, a query “Zhang Katz” on a university database may find the name “Zhang” occurring in a student tuple, and the name “Katz” in an instructor tuple, and a path through the advisor relation connecting the two tuples. Other paths, such as student “Zhang” taking a course taught by “Katz” may also be found in response to this query. Such queries may be used for ad hoc browsing and querying of data, when the user does not know the exact schema and does not wish to take the effort to write an SQL query defining what she is searching for. Indeed it is unreasonable to expect lay users to write queries in a structured query language, whereas keyword querying is quite natural.
Since queries are not fully defined, they may have many different types of answers, which must be ranked. A number of techniques have been proposed to rank answers in such a setting, based on the lengths of connecting paths, and on techniques for assigning directions and weights to edges. Techniques have also been proposed for assigning popularity ranks to tuples and XML elements, based
21.9 Directories and Categories 935
on links such as foreign key and IDREF links. See the bibliographical notes for more information on keyword searching of relational and XML data.